- Smartwatch sensors plus a machine learning algorithm successfully identified atrial fibrillation in people being treated for abnormal rhythm, but at lower accuracy when Afib was self-reported, a new study in JAMA Cardiology shows.
- Researchers from the University of California-San Francisco and Cardiogram analyzed data on 9,570 participants with Apple Watch enrolled in the Health eHeart study, 347 of whom self-reported Afib. Another group of 51 patients was undergoing cardioversion therapy. The raw sensor measurements — heart rate and step counts — were used to build a deep neural network, or algorithm.
- Among the 51 cardioversion patients, the technique detected Afib with 98% sensitivity and 90.2% specificity. Among self-reported participants in a larger group of 1,617 ambulatory individuals, sensitivity was 67.7% and specificity was 67.6%.
Wearables and other sensor technologies holds big potential to improve population health and remote monitoring of chronic conditions.
Evidence has been mixed on wearables but literature is growing around Apple Watch when used as a medical device for Afib detection.
In a recent Cleveland Clinic study, AliveCor’s KardiaBand app for Apple Watch accurately detected Afib versus normal sinus rhythm with 93% sensitivity and 84% specificity. Sensitivity increased to 99% when doctors reviewed the portable electrocardiogram recordings. A separate Mayo Clinic study showed the app can detect high potassium levels in the blood.
Roughly 34 million people worldwide have Afib, a leading cause of stroke. Because individuals often have no symptoms, early detection and monitoring are critical in preventing serious or life-threatening complications.
Deep neural networks are a form of machine learning algorithm particularly adept at pattern recognition. It has proven to be on a par with clinicians in classifying skin melanoma, ECG rhythm strip interpretation and diabetic retinopathy images and other anomalies, the researchers wrote.
While the deep neural network’s Afib classification shows promise, there were challenges detecting abnormal heart rates in ambulatory individuals moving about in natural environments.
There is also the possibility the results would not generalize to less tech-savvy individuals, the study notes, adding all of the participants already owned smartphones or had assistance from a care coordinator.
While previous research has shown that sensors can measure pulsatile blood flow, most have involved tightly controlled data, an accompanying editorial notes.
“[W]hat few other studies have done is to show prospective real-world validation, as these investigators have,” an accompanying editorial says. “The result is indeed humbling, indicating far more misclassification and lower positive predictive value compared with gold-standard clinician interpretation of 12-lead or smartphone-based limb-lead ECGs.”